Methods — STREAM v2.0 dashboard build

Marker genes

We compute per-cluster differential expression with Scanpy’s tl.rank_genes_groups using three statistics (wilcoxon, t-test, logreg). For each method we contrast the selected leiden cluster against the rest, keep genes with adjusted p < 0.05 (Bonferroni), and display the top hits by the method’s score (up to 50 per cluster). For readability, known aliases are appended to symbols in the table (display only).

Neighbor-aware “Top markers” (new)

To make marker calls more specific to local micro-populations, we compare each cluster against its K most-connected neighboring clusters on the KNN graph (default K = 3). We build a cluster-by-cluster connectivity matrix by summing cell-level graph weights (obsp['connectivities']), pick the top-K neighbors per cluster, and run rank_genes_groups (wilcoxon) with the target cluster as the group and the pooled neighbors as the reference. Results are filtered at adj. p ≤ 0.05 and ranked by |logFC|; we keep the top 50 per cluster. If a neighbor graph is missing, we compute PCA and pp.neighbors on the fly, or gracefully fall back to cluster-vs-rest if too few clusters exist.

Cell-type prediction (enrichment)

We score cell-type signatures with decoupler’s ULM method against a curated PanglaoDB marker resource (human entries, canonical markers only, sensitivity > 0.5, duplicates removed). ULM scores are computed per cell (dc.mt.ulm, tmin=3), then summarized per cluster using rankby_group (t-test with overestimated variance); only positive-stat entries are kept. The “Predicted annotation” tab shows (i) a UMAP colored by per-cell enrichment for the selected cell type and (ii) per-cluster score distributions. We list the top 5 candidate cell types per cluster as suggestions, not hard labels.

Transcription-factor programs

TF activity is estimated with decoupler ULM using the human CollecTRI regulon network. We compute per-cell TF scores, then rank TFs per cluster (same statistics as above) and retain the top 5 TFs as “program markers.” The TF tab shows a UMAP of the selected TF’s activity and a violin panel with cluster-wise distributions.

Co-expression & gene-set analysis

Co-expression: given Gene A and Gene B, cells are colored either (a) in bivariate mode using independent quantile thresholds for A and B (A-high, B-high, both-high, low) or (b) in ratio mode using A/B (or log₂(A/B)). Controls include per-gene clipping to upper quantiles, binarization, a global scale option, and an Otsu-based auto-threshold “Suggest” button for each gene. An auxiliary table lists cells with the most extreme A/B ratios.

Gene-set scoring: the “Signature” panel computes a per-cell mean Z-score across the pasted gene list (each gene is Z-scored across cells; the signature score is their mean). The resulting score is rendered on UMAP and also registered as a virtual gene so it can be reused elsewhere in the dashboard.

AI (A1) insights — optional

If enabled at build time (“Generate biology insights”), we produce cluster-wise narrative summaries with a large language model (Gemini family with automatic fallbacks). Only cluster-level summaries are provided to the model: cluster sizes; neighbor-aware top markers; method-specific marker tables; the per-cluster PanglaoDB enrichment ranks; and the top TFs. Raw counts or per-cell expression matrices are not sent. The model returns concise Markdown with a proposed label (± a refined subtype), confidence (High/Medium/Low), and supporting markers. Treat AI text as suggestive and validate against the Markers, Predicted annotation, and TF tabs.

Rendering notes: expression vectors are quantized for sparsity-aware transport and colored with a grey→rainbow palette; QC and composition views reflect the currently selected layer.